Skip to content

EleutherAGI/summarisation

Repository files navigation

Eleuther experiments in general intelligence

Learning from human feedback

Large language models learn a lot of very useful information about the world, we are experimenting with human preferences to steer the models

The task of summarisation acts as a good testing ground for comparing methods as OpenAI have set some baselines and released their datasets here.

Links

Download comparisons from OpenAI azcopy copy "https://openaipublic.blob.core.windows.net/summarize-from-feedback/dataset/*" . --recursive.

OpenAI's Dataset explorer

(not currently working) https://openaipublic.blob.core.windows.net/summarize-from-feedback/datasets/tldr_3_filtered and https://openaipublic.blob.core.windows.net/summarize-from-feedback/datasets/tldr_3_filtered_queries also host OpenAI's filtered verson of the dataset TL;DR dataset by Syed, Shahbaz, Voelske, Michael, Potthast, Martin, & Stein, Benno (2018). It is licensed under CC BY 4.0.

About

The Intermediate Goal of the project is to train a GPT like architecture to learn to summarise reddit posts from human preferences, as this has been done by OpenAI and provides a good benchmark to compare against. We will use this intermediate step as a way to lay the groundwork needed for on the fly learning using implicit models.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors